Active chemical sensing

Scope

A chemical sensor does not receive an odor, analyte identity, source, or hazard. It receives a time-dependent response produced jointly by source release, transport, reaction, surfaces, the receiver's path, the sampling action, the inlet and chamber, sensor chemistry, temperature, humidity, calibration, adaptation, ageing, contamination, and previous exposure. In a turbulent plume, even the material reaching the receiver arrives as intermittent whiffs and blanks rather than a smooth pointer to its source.

This chapter turns the olfaction, chemical sensing, and plume-tracking audit into readable architecture. The detailed definitions live in operator-qualified chemical-sensing mathematics, and Fixture F-011 tests the architecture across fourteen hostile tracks. The editable diagram is kept in operator-qualified-active-chemical-sensing.mmd.

The chapter connects existing project components rather than adding another principle or candidate:

1. sensorimotor grounding, because sniffing, pumping, orientation, locomotion, and receiver geometry change the evidence;
2. operator-qualified sensing, because every chemical result remains conditional on a physical forward operator;
3. sparse predictive compute, because temporal events and sparse representations earn efficiency credit only through total task work and measured energy;
4. memory and consolidation, because fast adaptation, learned associations, slow calibration, drift, and maintenance occupy different state and update timescales;
5. reliability under mission profiles, because humidity, contamination, ageing, poisoning, replacement, and out-of-support operation change the device rather than merely the data; and
6. the energy model, because motion, pumps, heaters, preconcentration, chromatography, vacuum/ionization, calibration gases, consumables, maintenance, human work, and embodied devices remain inside the service boundary.

The intended output is a calibrated decision, qualified retained observation, safe action, or abstention. Presence, molecular identity, perceptual odor identity, concentration, mixture composition, direction, source position, source attribution, intensity, valence, hazard, exposure, and absorbed dose are different outcomes. One may help predict another; none may silently replace it.

The evidence range for this chapter is C-1152–C-1203: 50 claims are established within their stated experiments or authoritative methods, C-1188 is plausible, and C-1192 is disputed. Those statuses qualify the claim boundaries; they are not votes for the architecture.

Biological observation

Receptor populations provide coverage, not self-describing identities

Mammalian olfaction begins with a large receptor family. Within the receptor panel and odorants studied, individual receptors responded to multiple odorants, individual odorants recruited multiple receptors, and the population pattern changed with concentration (C-1152–C-1154). This supports a distributed measurement basis. It does not supply a universal odor code, open-world chemical coverage, or concentration-independent token.

Chemosensation itself is not one architecture. Manipulations of mammalian sweet and umami pathways provide a scoped dedicated-cell counterexample (C-1155). A receptor-like array must therefore state which chemicals and concentrations it can distinguish, where responses overlap, and which outcomes require another channel. “Combinatorial,” “labelled-line,” or “olfactory” is a description of evidence organization, not an implementation credit.

Bulb and cortex transform concentration, gain, and timing

Divisive normalization in the studied fly antennal lobe scaled projection-neuron responses with pooled receptor activity. Mouse olfactory-bulb and piriform measurements show transformations that can make identity representations more tolerant to concentration, while still preserving useful intensity or concentration-change information in other activity (C-1156–C-1162).

Three constraints follow:

1. concentration tolerance is an outcome to measure, not an invariant built into an anatomical label;
2. suppressing absolute level can damage leak, dose, or safety tasks even when identity improves; and
3. pooled inhibition and recurrence must compete with robust scaling, explicit gain-state estimation, concentration-conditioned inference, and ordinary recurrent models at equal latency and energy.

Receptor adaptation includes causal calcium-dependent feedback, receptor current and spike output can span different concentration ranges, and habituation depends on duration, interval, and odor similarity (C-1163–C-1166). Receptor adaptation, behavioral habituation, short-term sensor recovery, calibration drift, and irreversible poisoning must remain separate states.

Sniffing is part of the observation operator

Rats can alter sniff rate rapidly during discrimination; changing sniffing changes the peripheral-to-bulb filter; early inhalation-locked activity can carry information quickly; and mice can use millisecond-scale sniff-phase differences under scoped protocols (C-1167–C-1171). That timing cannot recover fluctuations already removed by tubing, a chamber, or a slow sensor. End-to-end bandwidth, capture and receipt time, and the realized sampling waveform set the usable support.

Spatial acquisition is also regime-dependent. Serial sampling sufficed in one mouse gradient task, while bilateral temporal correlation supplied odor-motion information in a fly preparation (C-1172–C-1173). Neither result establishes universal stereo or universal serial sampling. Body size, receptor spacing, movement, wind, range, plume intermittency, and sensor bandwidth decide which comparison is informative.

Turbulent plumes turn localization into inference under intermittent evidence

Theory tested against simulation, laboratory, and field observations describes structured whiff and blank statistics rather than a smooth instantaneous gradient. Fine plume structure requires fast ground truth; field gaps change with environment; and several intuitive burst summaries converge too slowly or vary too weakly to guide short-horizon search reliably (C-1174–C-1177).

Animals demonstrate multiple bounded strategies. Moths can surge after odor contact and cast after loss. Walking flies use distinct odor-ON, odor-OFF, and wind transforms, and in irregular plumes their stochastic turns and walk/stop decisions depend on encounter timing rather than continuous steering (C-1178–C-1180). Contact-correlated slowing in mice is an observation, not proof that slowing is optimal (C-1181).

Robotics already supplies strong nulls. Infotaxis, gas-plus-wind localization, transient processing for slow metal-oxide sensors, reactive off-zigzag search, geometry-aware modular policies, and particle-filter source belief have all been demonstrated in scoped settings (C-1182–C-1187). Compact temporal-memory reinforcement learning is only plausible for the studied simulated plumes until unchanged embodied transfer is shown (C-1188).

Sparse piriform activity is not a hardware or energy conclusion

Piriform odor responses can be sparse and distributed rather than neatly topographic; in one rat task, burst-count population information outperformed some precise-pattern accounts (C-1189–C-1191). The degree of sparsity is disputed as a universal characterization because it changes with concentration and protocol (C-1192). Longitudinal piriform ensembles can also drift while behavior remains stable (C-1193).

These observations motivate tests for sparse routing, event memory, and remappable readouts. They do not establish fixed semantic neuron addresses, low memory traffic, low organism energy, or a benefit over pruning, compression, low precision, sparse convolution, or dense execution on suitable hardware.

Association, valence, identity, and hazard are separable

Rapid reward-category coding appeared in olfactory tubercle within minutes in one task, while posterior piriform lacked the same explicit code even after overtraining. Arbitrary piriform ensembles could acquire opposite valence under different reinforcement, and innate aversion could be disrupted while learned detection or avoidance remained (C-1194–C-1196). Cortical-amygdala pathways causally contributed to scoped innate odor behavior (C-1197).

The architectural record must therefore preserve odor evidence, reinforcement, context, action, feedback, acquisition time, retention, transfer, reversal, and readout/module identity. Chemical identity, odor category, innate choice, learned choice, pleasantness, irritation, toxicity, external exposure, and hazard remain separately scored.

Mixtures and instruments expose the same ambiguity problem

Animals can learn a target in variable mixtures, but performance degrades with background count and overlap; chemically similar maskers can raise detection thresholds more in the tested regime (C-1198–C-1199). Cross-reactive artificial arrays are already an established sensing baseline, not a novel consequence of receptor analogy (C-1200).

Multi-year metal-oxide sensor data demonstrate drift. Humidity, stability, selectivity, and poisoning belong to the operator state, and validated analytical/safety workflows preserve sampling, calibration, recovery, identification support, exposure units, and uncertainty (C-1201–C-1203). A high closed-panel classifier score cannot turn an unsupported mixture into an identified chemical, a library match into source attribution, or odor detection into safety.

Proposed AI translation

Preserve the whole chemical episode

For episode $e$, preserve

where:

• $S_e$ records source identity, mixture, release in moles per second, temperature in kelvins, geometry in metres, phase, and motion in metres per second;
• $X_e$ records domain, boundaries, surfaces, airflow in metres per second, pressure in pascals, relative humidity as a dimensionless fraction, temperature, turbulence, reaction, sorption, and chemical background;
• $A_e$ records commanded and realized motion, orientation, sniff/pump flow in cubic metres per second, heater power in watts, valve, preconcentration, purge, query, confirmation, stopping, and abstention;
• $R_e$ records receiver/body, bilateral or array geometry, pose, inlet, tubing, chamber, pump, heater, sensor, saturation, health, and feasible authority;
• $O_e$ is the versioned observation operator: response/recovery, cross- sensitivity, nonlinearity, hysteresis, support, clock, latency, quantization, preprocessing, missingness, and selection;
• $K_e$ records reference-gas composition and uncertainty, blanks, zero/span, flow, device and batch, compensation, age, drift, poisoning, maintenance, calibration validity, and traceability;
• $H_e$ records prior exposure, adaptation, habituation, contamination, cleaning, training, reinforcement, feedback, previous actions, and readout remapping with timestamps;
• $T_e$ declares the literal target, deadline in seconds, loss/utility, abstention policy, exposure rule, and safety constraint;
• $U_e$ names the independent unit: sample, stock, injection, device, batch, day, source, plume realization, site, body, animal/subject, or model seed; and
• $B_e$ is the componentwise ceiling in evidence, labels, standards, channels, actions, metres, seconds, bytes, searches, person-hours, joules, consumables, exposure, replacements, embodied devices, and opportunity.

This is the chemical instantiation of the versioned observation contract. The endogenous-observation candidate owns the coupling between acquisition action and future evidence; the latency-qualified authority envelope owns the action restriction when evidence is slow, stale, saturated, miscalibrated, or poisoned.

Model transport before interpreting the sensor

For analyte $i$, a minimum transport model is

where amount concentration $c_i$ is in moles per cubic metre, position $\mathbf x$ is in metres, time $t$ is in seconds, velocity $\mathbf u$ is in metres per second, diffusivity or declared effective dispersion $D_i$ is in square metres per second, relative humidity $H_r$ is dimensionless, and reaction/loss/phase-transfer $R_i$ and volumetric source $q_i$ are in moles per cubic metre per second. Every term has units of moles per cubic metre per second. Boundaries, buoyancy, droplets, thermal stratification, deposition, and unresolved turbulent fluxes remain explicit when they affect the task.

For channel $m$ sampled at $t_n$, the measured trace is

where channel output $y_{m,n}$ and error $\epsilon_{m,n}$ use the calibrated sensor unit, causal response kernel $h_{m,i,v}$ is in reciprocal seconds, delay $\tau$ is in seconds, receiver path $\mathbf x_r$ is in metres, version $v$ is dimensionless, and $\mathbf z_n$ contains flow, heater, temperature, humidity, interferents, adaptation, age, drift, saturation, and poisoning. A static feature vector is permitted only after this dynamic operator has been tested or shown irrelevant inside the declared support.

Represent non-identifiability instead of forcing a label

Let $G_v$ be the calibrated mixture-to-sensor forward operator and $\mathcal S_c$ the supported set of nonnegative composition vectors in moles per cubic metre. For observation $\mathbf y$, retain

where error covariance $\Sigma_y$ is in squared sensor-output units, the Mahalanobis norm and tolerance $\varepsilon_y$ are dimensionless, and $\mathcal N_v$ is the observation-equivalent composition set. If materially different identities, concentrations, exposures, or hazard states remain in that set, the result is ambiguous. The system can acquire another measurement, request analytical confirmation, retain alternatives, or abstain; it cannot convert a prior-selected label into new chemical evidence.

This state links to reset-coupled staged verification: an inexpensive array may screen, but escalation to GC--MS, PTR/SIFT--MS, IMS/FAIMS, or another qualified method must add conditionally useful evidence after sampling, standards, turnaround, analyst time, consumables, exposure, and energy are charged. The analytical method is itself an operator with blanks, recovery, retention, deconvolution, library support, calibration, and uncertainty—not an oracle.

Keep a literal outcome firewall

Relative gradients are valid only for the field transformation tested

Multiplying an entire positive concentration field by a constant preserves its shape and every local ratio, but adding a background, changing transport, or clipping a receptor does not. The E. coli evidence links response rescaling to search only inside finite concentration regimes (C-1542); the social-amoeba evidence adds a density-qualified secrete-and-sense boundary rather than generic density-independence (C-1546).

An active-search policy therefore records which field transformation it is expected to ignore. It is challenged with same-ratio/different-difference and same-difference/different-ratio fields, additive backgrounds, unseen source strengths, saturation, transport change and values near zero. A relative channel may guide search only while its support gate remains valid; an absolute-critical exposure or load stays on a calibrated absolute channel. The full mathematical and protocol boundary is kept in Interface-qualified scale symmetry and Fixture F-026.

Use action to change observability, not to obtain a free second dataset

At decision time $t$, choose

where $\mathcal H_t$ is causally received observation/action history, $\widehat{\mathbf c}_t$ is concentration/mixture belief in moles per cubic metre, $\widehat{\mathbf s}_t$ is source belief with position in metres and release in moles per second, $\widehat O_t$ is operator/condition state, $\widehat U_t$ is uncertainty, $\mathcal A_t^{\mathrm{safe}}$ is the independently constrained action set, and $\mathbf B_t$ is remaining budget in its component units.

The action may move or orient the body, change bilateral spacing, sniff or pump, change heater or valve state, purge, resample, request confirmation, stop, or abstain. Its causal value must be tested against fixed, random, replayed, and dose-matched acquisition. Equal wall time is insufficient if one method inhales or pumps more material, experiences more whiffs, travels farther, uses more energy, or accepts more exposure.

For plume search, the system retains a joint belief

not one gradient arrow. Surge--cast/off-zigzag rules, wind-only anemotaxis, particle-filter belief control, infotaxis, finite-state search, POMDP/MPC/value of information, and matched-memory reinforcement learning remain mandatory nulls. Success, false source declarations, location error, posterior coverage, path in metres, time in seconds, collisions, exposure, and joules are reported separately.

Maintain fast response and slow condition as different states

Use at least two state transitions:

where within-episode state $\mathbf r_n$ includes response, adaptation, heater, and recovery; between-episode state $\mathbf d_e$ includes calibration, baseline/gain drift, contamination, ageing, and poisoning; concentration $\mathbf c_n$ is in moles per cubic metre; cumulative stress/exposure $\mathcal E_e$ retains its physical units; and maintenance action $m_e$ records purge, cleaning, recalibration, repair, or replacement. A return to baseline does not prove restored selectivity or calibration. A task residual cannot by itself distinguish environmental change from device change.

The graded assurance envelope binds calibration and condition evidence to the exact operator version. The reversible physical-skill candidate receives credit for coatings, inlets, chambers, filters, heaters, or other physical transforms only after cross-sensitivity, reset, poisoning, replacement, fallback, and fabrication burden are measured.

Preserve evidence for recalibration and future interpretation

Raw traces, calibration/operator history, standards, sample lineage, analytical evidence, and retained physical samples have different reconstruction value. Contract-preserving compaction may replace them only for registered future queries; value- and reconstructability-aware tiering must survive hidden recalibration, changed-library, changed-exposure-rule, and poisoning-investigation queries without future-label leakage.

One closed sensing-and-action contract

flowchart TB
    source["Source and release<br/>identity · mixture · rate · geometry · motion"] --> transport["Transport and transformation<br/>advection · turbulence · diffusion · reaction · sorption"]
    environment["Environment state<br/>boundaries · wind · temperature · humidity · pressure"] --> transport
    transport --> field["Intermittent chemical field<br/>whiffs · blanks · concentration · composition"]
    action["Acquisition action<br/>sniff/pump · move · orient · heat · valve · purge"] --> receiver["Receiver and sampling path<br/>body · inlet · tubing · flow · chamber · aperture"]
    field --> receiver
    receiver --> operator["Versioned observation operator<br/>response/recovery · cross-sensitivity · saturation · support"]
    condition["Operator condition<br/>calibration · adaptation · age · humidity · drift · poisoning"] --> operator
    operator --> observation["Causally received trace/events<br/>values · timestamps · missingness · uncertainty"]
    observation --> inference["Calibrated inference<br/>detect · identify · quantify · separate · localize · abstain"]
    history["Causal history<br/>prior exposure · actions · learning · maintenance"] --> condition
    history --> inference
    inference --> firewall["Outcome firewall<br/>presence · identity · concentration · mixture<br/>source · valence · exposure · hazard"]
    inference --> decision["Decision<br/>act · resample · move · confirm · stop · abstain"]
    decision --> action
    decision --> safety["Independent safety envelope<br/>exposure limits · authority · fail-safe action"]
    safety --> action
    analytical["Analytical confirmation nulls<br/>GC–MS · PTR/SIFT–MS · IMS/FAIMS · standards"] --> confirmation["Qualified confirmation<br/>blanks · recovery · retention/spectral evidence"]
    decision --> confirmation
    confirmation --> firewall
    nulls["Mature null stack<br/>dynamic calibration · chemometrics · state estimation<br/>surge–cast · particle filter · infotaxis · POMDP/MPC/VOI"] --> compare{"Equal evidence · action · exposure · lifecycle budget"}
    firewall --> compare
    ledger["Complete ledger<br/>samples · standards · time · person-hours<br/>motion/pump/heater/analysis · operational + embodied joules"] --> compare
    compare --> retain["Retain only literal track residual"]
    compare --> retire["Retire composition<br/>preserve chemical observation contract"]

Editable source: operator-qualified-active-chemical-sensing.mmd.

Efficiency mechanism

The architecture permits six distinct efficiency mechanisms. Each remains a hypothesis until it improves a protected outcome under F-011's matched budget.

Sparse or event-driven computation is not a seventh saving until its total physical cost is lower. A top-$k$ or thresholded representation can reduce arithmetic while adding normalization, sorting, indices, irregular memory traffic, routing, remapping, idle hardware, missed-event risk, and maintenance. The relevant numerator is accepted task service, not active-unit count.

Lifecycle energy for method $q$ over one accepted service interval is

where every term is energy in joules and covers data acquisition, training, receiver motion, pumping, heating, sensing, analytical separation, ionization/vacuum, inference, communication, storage, calibration, maintenance, facility overhead, and amortized embodied hardware. Carrier and calibration gases, sorbents, columns, dopants, filters, cleaning agents, samples, emissions, and disposal remain additionally reported in their native physical or lifecycle units.

Human work is

where every term is in person-hours and roles are separated. An apparent energy gain is rejected if it moves work into sample preparation, calibration, chemical analysis, safety review, data curation, cleaning, or repair without counting it.

Evidence status

The totals are exactly 50 established, one plausible, and one disputed claim. The established status applies only to the cited biological preparation, behavior, instrument, dataset, method, or authoritative standard. It does not establish that the project composition improves an engineering frontier.

The complete mature null is a composition, not a token baseline:

1. traceable sampling, standards, blanks, duplicates, recovery, flow and calibration;
2. GC--MS/GC--FID/PID and GC×GC where justified, retention indices, authentic standards, PTR/SIFT--MS, IMS/FAIMS, electrochemical/PID, and targeted spectroscopy under their support;
3. dynamic system identification, response/recovery modelling, deconvolution, filtering, robust scaling and temperature/humidity compensation;
4. PCA/PLS, LDA/QDA, calibrated regression/classification, SVMs, trees, ensembles, neural models, open-set detection, conformal/selective prediction, mixture models, domain adaptation and abstention;
5. measured/validated flow models, Kalman/particle filtering, Gaussian-process plume inference, observability and posterior calibration;
6. correlated random walk, gradient and wind baselines, surge--cast, off-zigzag, infotaxis, particle-belief control, finite-state search, POMDP/dual control/MPC/value of information and matched-memory RL; and
7. detector health, poisoning/out-of-support alarms, staged verification, exposure constraints, independent authority, fallback, maintenance, sample lineage, human work, consumables, and lifecycle accounting.

F-011 compares against that full stack. A weak static classifier, uncalibrated e-nose, single gradient controller, or instrument name is not the baseline.

Speculative extensions

The following are experiment-generating compositions. None is promoted by this chapter.

Action-conditioned identifiability

Use the current observation-equivalent set $\mathcal N_v(\mathbf y)$ to select the cheapest safe action expected to separate decision-relevant alternatives. The action might change path, wind-relative orientation, bilateral geometry, flow, heater state, temporal support, or analytical method. This joins Candidate 007 with Candidate 014. It survives only if explicit Bayesian design, value of information, POMDP/dual control, and ordinary staged testing cannot reach the same calibrated decision frontier.

Dual identity--concentration state

Maintain shared evidence with separately protected readouts for chemical/odor identity, absolute concentration, concentration change, and operator condition. Normalization or recurrence may stabilize the identity readout while the other paths preserve dose and condition. A useful implementation must beat concentration-conditioned generative models and explicit gain-state estimators; it is rejected when one task improves by destroying another.

Remappable sparse population memory

Treat sensor/receptor channels and sparse learned units as replaceable evidence contributors rather than permanent semantic addresses. A readout-maintenance layer would detect drift, remap channels, preserve uncertainty, and request labels or calibration selectively. It is worth retaining only if future-time performance improves after update writes, labels, monitoring, downtime, memory traffic, replacement, and energy are charged. Standard recalibration, domain adaptation, ensemble remapping, pruning, compression, and dense low-precision execution remain the nulls.

Operator-matched event front end

Co-design inlet, chamber, sensor physics, heater/pump action, deconvolution, and event thresholds so the retained trace preserves task-bearing whiff/blank and transient information at lower traffic. This is a scoped extension of Candidate 006, not a claim that physical or event-driven sensing is intrinsically efficient. It must transfer across hardware, humidity, drift and plume timescale and must beat a calibrated dynamic model on the same device.

Qualified screen--confirm--retain loop

Compose an inexpensive cross-reactive screen, calibrated abstention, conditional analytical confirmation, and query-aware retention. The screen can provisionally act only inside the latency-qualified authority envelope; Candidate 010 owns escalation; Candidates 017 and 018 own retained evidence. The composition is rejected if an ordinary calibrated cascade or always-confirm policy matches protected risk, latency, and total cost.

Regime-switching plume search

Use calibrated transport, sensor-condition and source beliefs to switch among reactive ON/OFF behavior, wind-relative movement, local search, belief-driven exploration, confirmation, and safe withdrawal. The controller must expose the regime evidence that authorized the switch and must abstain when operator or wind evidence is invalid. It is rejected if one finite-state controller, particle-belief policy, infotaxis, POMDP/MPC, or matched-memory learner reaches the same held-out source-search frontier.

Failure modes

Physics and operator failures

1. Smooth-gradient fiction: instantaneous concentration is treated as a stable source direction despite intermittent transport.
2. Static-vector fiction: inlet, chamber, response, recovery, hysteresis, saturation, humidity, and prior exposure are discarded before inference.
3. Bandwidth invention: millisecond or event information is claimed after the physical transport or sensor has filtered it away.
4. Mixture over-identification: a single label is emitted while materially different compositions, concentrations, exposures, or hazards remain observation-equivalent.
5. Conversion error: parts per million are converted to mass concentration without molar mass, temperature, pressure, and fraction definition.
6. Transport/operator confounding: policy performance is credited to the learner although one arm received better wind, likelihood, sensor dynamics, field truth, calibration, or source prior.

Representation and learning failures

7. Label-as-mechanism: “receptor-like,” “bulb,” “piriform,” “sparse,” “temporal,” or “neuromorphic” replaces a causal ablation and literal endpoint.
8. Coverage-as-decoder gain: more sensor chemistry, area, concentration, standards, or SNR is attributed to architecture.
9. Concentration erasure: identity appears invariant because the system discarded information required for leak, exposure, or safety decisions.
10. Stable-address assumption: drifting or replaced sensor/representation units retain fixed semantic addresses without remapping cost.
11. Event-count efficiency: fewer active events are reported without bytes, memory traffic, routing, decoding, idle hardware, missed hazards, and lifecycle joules.
12. Association collapse: endpoint accuracy substitutes for acquisition curve, reinforcement/context, retention, transfer, reversal, and selective representation/readout intervention.
13. Valence collapse: innate choice, learned choice, pleasantness, irritation, toxicity, exposure, and hazard are merged into one score.

Evaluation, reliability, and safety failures

14. Temporal/batch leakage: random rows or adjacent windows share stock, dilution, sample, plume seed, sensor, device, batch, calibration, day, site, subject, or analytical run across splits.
15. Simulation inverse crime: training and evaluation share CFD mesh, response kernel, source schedule, random seed, or post-test retuning.
16. No-source omission: every episode contains a source, so unconditional declarations look successful and false reassurance stays invisible.
17. Compensation without condition detection: expected drift is corrected while humidity, contamination, poisoning, replacement, or out-of-support state remains undetected.
18. Instrument-as-oracle: GC--MS or another method is credited without sample lineage, blanks, recovery, breakthrough/carryover, separation, retention/spectral support, standards, library scope, and uncertainty.
19. Odor-as-safety: detection threshold, intensity, preference, or aversion substitutes for a chemical-, route-, population-, endpoint-, and averaging- time-specific exposure rule.
20. Self-certified authority: the same uncertain sensor/inference path defines its own safety envelope and fallback.
21. Free active sensing: an adaptive method samples more material, sees more whiffs, moves farther, waits longer, consumes more pump/heater energy, or accepts more exposure than its baseline.
22. Incomplete lifecycle boundary: analytical preparation, standards, calibration gases, consumables, motion, pumps, heaters, chromatography, vacuum/ionization, facility power, cleaning, replacement, human work, embodied devices, emissions, and disposal disappear from the ledger.

Any failure that creates the reported advantage retires the architectural claim for that track. The operator record can remain useful even when the proposed mechanism does not.

Measurable predictions

Fixture F-011 implements these predictions with frozen splits, common operator/action budgets, causal ablations, source-off trials, prospective device/time tests, and full resource accounting.

For every track, report literal outcomes, calibrated uncertainty, independent unit, failures, abstentions, unused budget, and the complete cost vector. A residual must replicate across at least two target chemical families, interferent/matrix families, concentration and source/plume regimes, sensor chemistries, manufacture batches, operator/calibration versions, future times, sites, model families, and hardware classes. Active tracks additionally require unseen source positions, plume seeds, bodies, action limits, and paired counterfactual seeds.

If the complete mature stack matches the composition, if no selective ablation isolates value, or if the gain disappears after calibration, analytical work, exposure, maintenance, human effort, consumables, facility and lifecycle energy are charged, retire the architectural residual. Keep the chemical observation contract and the negative result; create no new principle or candidate.